Erik Kusch, PhD
Code Coffee @ CICERO
2024-11-14, Forskningsparken
Statistical Downscaling &
Geostatistical Interpolation
of Climate and Weather Data
Uses-Cases, Specifications, Pitfalls
3
Temporal
…
Spatial
Dynamical Mechanistic Statistical
Statistical
Downscaling
Delta Change Weather
Classification Regression …
Geostatistical
Interpolation
Inverse
Distance
Weighting
(Thin-Plate)
Splines
Trend
Surfaces Kriging …
Hybrid Machine
Learning …
Downscaling Methodology Overview
Dynamical
Use high-resolution regional climate
modes (RCMs) to simulate detailed
local climate based on global climate
model (GCM) data, capturing complex
terrain and weather patterns.
Mechanistic
Statistical
Hybrid Solutions
Machine Learning
4
Downscaling Approaches
Dynamical
Mechanistic
Statistical
Employ statistical relationships
between large-scale climate variables
and local observations to project finer-
scale climate outcomes.
Hybrid Solutions
Machine Learning
6
Downscaling Approaches
ISSN 1039–7205; Copyright © Western Australian Agriculture Authority, 2011
Dynamical
Mechanistic
Statistical
Hybrid Solutions
Combine statistical and dynamical
downscaling methods to leverage both
physical modeling and empirical data
relationships, improving accuracy and
efficiency for local climate predictions.
Machine Learning
7
https://doi.org/10.1175/MWR-D-19-0196.1
Downscaling Approaches
Dynamical
Mechanistic
Statistical
Hybrid Solutions
Machine Learning
Use algorithms to learn patterns from
historical climate data, enabling high-
resolution climate projections by
predicting local climate details from
large-scale climate model outputs.
8
Downscaling Approaches
The ARC Centre of Excellence for Climate Extremes | Using Machine Learning to Cut the Cost of Downscaling Global
Climate Models - The ARC Centre of Excellence for Climate Extremes
9
Comparing Downscaling Approaches
Downscaling
Approach Accuracy
Computational Cost
Expertise Required Other Important Criteria
Dynamical
High accuracy in capturing physical processes,
especially in complex terrains
Very high
Advanced
Provides detailed physical representations,
sensitive to model assumptions
Mechanistic
Moderate to high accuracy; dependent on process
representation quality
Moderate to high
Moderate to advanced
Best for studies focusing on specific physical
processes at local scales
Statistical
Moderate accuracy; limited by statistical model
assumptions
Low
Moderate
Fast, resource
-
efficient, but may lack precision in
extreme events
Hybrid
Potentially high, combining benefits of dynamical and
statistical approaches
High
Advanced
Flexible, leveraging strengths of multiple
approaches, but complex to set up
Machine Learning
Moderate to high accuracy, depending on model and
training data
Moderate to high
Moderate to advanced
Accuracy depends on data quality; may generalize
poorly outside trained range
10
Temporal
…
Spatial
Dynamical Mechanistic Statistical
Statistical
Downscaling
Delta Change Weather
Classification Regression …
Geostatistical
Interpolation
Inverse
Distance
Weighting
(Thin-Plate)
Splines
Trend
Surfaces Kriging …
Hybrid Machine
Learning …
Downscaling Methodology Overview
Relate coarse-scale climate data to fine-scale climate data.
Purpose:
Generate fine-scale local climate projections from coarse-resolution global
climate model (GCM) or regional climate model (RCM) outputs.
Methodology:
Establish statistical relationships between large-scale predictors and local
climate observations.
Relate coarse-scale data to location in space and fine-scale covariates.
Purpose:
Estimate values at unsampled locations based on known observations,
creating continuous spatial datasets.
Methodology:
Utilize spatial statistical techniques that leverage spatial autocorrelation of
observed data points.
11
Temporal
…
Spatial
Dynamical Mechanistic Statistical
Statistical
Downscaling
Delta Change Weather
Classification Regression …
Geostatistical
Interpolation
Inverse
Distance
Weighting
(Thin-Plate)
Splines
Trend
Surfaces Kriging …
Hybrid Machine
Learning …
Downscaling Methodology Overview
Delta Change
Apply projected climate changes (or
"deltas") from GCMs to local historical
data: (1) Calculate changes in climate
variables GCM by comparing future
and baseline periods, (2) add these
calculated changes to local observed
data to create downscaled data.
Weather Classification
Regression
12
Statistical Downscaling Methods
http://dx.doi.org/10.3390/su12020477
Delta Change
Weather Classification
Grouping large-scale weather patterns
into classes or "weather types." Match
historical observations with these
classes to identify typical local
responses to each weather pattern.
Project relationships to downscale data.
Regression
13
Statistical Downscaling Methods
https://cpaess.ucar.edu/sites/default/files/meetings/2010/documents/stickel_lorna.pdf
Delta Change
Weather Classification
Regression
Establish statistical relationships
between variables from GCMs or RCMs
and local observations. Regress
observed variability at local scales
against the variability in climate model
outputs to derive equations that
capture these relationships. These
equations are then applied to create
downscaled data.
14
Statistical Downscaling Methods
https://link.springer.com/article/10.1007%2Fs12145-021-00669-4
CS
Down
scale
15
Temporal
…
Spatial
Dynamical Mechanistic Statistical
Statistical
Downscaling
Delta Change Weather
Classification Regression …
Geostatistical
Interpolation
Inverse
Distance
Weighting
(Thin-Plate)
Splines
Trend
Surfaces Kriging …
Hybrid Machine
Learning …
Downscaling Methodology Overview
Inverse Distance Weighting (IDW)
Estimates values based on the
distance of nearby known points,
with closer points having more
influence on the estimate.
(Thin-Plate) Splines
Trend Surface Analysis
Kriging
16
Geostatistical Interpolation Methods
https://gisgeography.com/inverse-distance-weighting-idw-interpolation/
Spat
Stat
Inverse Distance Weighting (IDW)
(Thin-Plate) Splines
Trend Surface Analysis
Kriging
Estimate unknown values based on
the spatial correlation of nearby
observed data points and, if desired,
covariate data. Weighs observations,
with closer points given more
influence.
19
Geostatistical Interpolation Methods
https://gisgeography.com/kriging-interpolation-prediction/
I developed and maintain this!
Use known relationships were
possible within your domain
(e.g., lapse rates)
20
Statistical Downscaling & Geostatistical Interpolation Considerations
Statistical Relationships
Ensure that relationships are
consistent at time-scale of your
study
Time-Scales
You can use variables you
already downscaled as
covariates
There needs to be substantial
variation within your data in
your domain
Co-Downscaling
Data Variation
Some variables depend on
dynamic covariates, others on
static covariates
Choice of Covariates
Don't do it Daily Weekly Monthly Annual
Use with caution Surface air temperature
All clear Soil moisture 0.5m and below
Soil moisture 0-0.5m
Radiation variables
Wind speed
Don't do it Daily Weekly Monthly Annual
Use with caution Surface air temperature
All clear Soil moisture 0.5m and below
Soil moisture 0-0.5m
Radiation variables
Wind speed
Statistical interpolation comes
with uncertainty –investigate
this and, when possible,
propagate it.
Uncertainty Propagation
21
Temporal
…
Spatial
Dynamical Mechanistic Statistical
Statistical
Downscaling
Delta Change Weather
Classification Regression …
Geostatistical
Interpolation
Inverse
Distance
Weighting
(Thin-Plate)
Splines
Trend
Surfaces Kriging …
Hybrid Machine
Learning …
Downscaling Methodology Overview
22
•Our different use-cases and data
requirements necessitate use of
different downscaling approaches
•Hurdle to carrying out downscaling
can be daunting:
ïƒ Toolbox collection of code routines to
downscale data at CICERO
ïƒ Downscaling evaluation routines
ïƒ Establishing best-practices for
statistical downscaling routines in our
projects
Statistical Downscaling & Geostatistical Interpolation and CICERO
Different strokes for different folks
Clim
Hub I propose to build this at CICERO!
Thank you
Erik Kusch
www.cicero.oslo.no
@CICERO_klima
